Enhancing Type 2 Diabetes Prediction through Transfer Learning: A Framework for Utilizing Unpaired Clinical and Genetic Data

The prevalence of type 2 diabetes mellitus (T2DM) in Korea has risen in recent years, yet many cases remain undiagnosed. Advanced artificial intelligence (AI) models using multi-modal data have shown promise in disease prediction, but two major challenges persist: the scarcity of samples containing all desired data modalities and class imbalance in T2DM datasets. We propose a novel transfer learning framework to predict T2DM onset within five years, using two Korean cohorts (KoGES and SNUH). To utilize unpaired multi-modal data, our approach transfers knowledge between clinical and genetic domains, leveraging unpaired clinical data alongside paired data. We also address class imbalance by applying a positively weighted binary cross-entropy (BCE) loss and a weighted random sampler (WRS). The transfer learning framework improved T2DM prediction performance. Using WRS and weighted BCE loss increased the model’s balanced accuracy and AUC (achieving test AUC 0.8441). Furthermore, combining transfer learning with intermediate data fusion yielded even higher performance (test AUC 0.8715). These enhancements were achieved despite limited paired multi-modal samples. Our framework effectively handles scarce paired data and class imbalance, leading to improved T2DM risk prediction. This approach can be adapted to other medical prediction tasks and integrated with additional data modalities, potentially aiding earlier diagnosis and better disease management in clinical settings.